WO2012140147A2

WO2012140147A2 - Processing ultrasound images

Info

Publication number: WO2012140147A2
Application number: PCT/EP2012/056691
Authority: WO
Inventors: Kevin Mcguinness; Sarah HUGHES; Niall MOYNA; Noel O'connor
Original assignee: Dublin City University
Priority date: 2011-04-12
Filing date: 2012-04-12
Publication date: 2012-10-18
Also published as: GB201106184D0; GB2490477A; WO2012140147A3

Abstract

A key indicator in the onset of Cardiovascular Disease (CVD) is abnormalities in the inner most layer of the arterial wall, i.e. the endothelial cell layer. Dysfunction of this endothelium is one of the earliest events in the development of CVD. Disclosed herein are inventive methods for processing an ultrasound image of a vascular tissue and an inventive method of segmenting the tissue for detecting sub abnormalities. The processing method entails: segmenting the ultrasound image into regions corresponding to vascular tissue lumen and vascular tissue; determining the diameter of the vascular tissue lumen at a plurality of points; determining an inter-percentile range of the plurality of diameters; and calculating a statistic of centre of the plurality of diameters from the inter-percentile range of the plurality of diameters.

Description

Title

Processing Ultrasound Images

Field of the Invention

[0001] The present invention relates to a method of processing ultrasound images and a novel method of segmenting the ultrasound images prior to processing thereof. The invention further discloses combining the method of processing with the method of segmenting as a tool in the detection of cardiovascular disease (CVD). Also disclosed are computer programs for implementing the methods of the invention.

Background to the Invention

[0002] Heart disease is a major contributor to mortality rates in the developed world.

Cardiovascular disease (CVD) relates to a group of disorders of the heart and the vasculature and includes coronary heart disease, congestive heart failure, peripheral arterial disease stroke and congenital heart defects.

[0003] Although many treatments for CVD exist, the World Health Organisation estimates that approximately 17.1 million people die of CVD each year. As with most life threatening conditions, early detection is key in the treatment and therapy of CVD.

[0004] A key indicator in the onset of CVD is abnormalities in the inner most layer of the arterial wall, i.e. the endothelial cell layer. Dysfunction of this endothelium is one of the earliest events in the development of CVD. In particular, the extent of vessel dilation such as arterial dilation (e.g. brachial artery dilation) after transient blockage (commonly referred to as brachial artery reactivity) and the thickness of the walls of the carotid arteries (commonly referred to as carotid intima media thickness) are indicators of endothelial dysfunction and CVD respectively.

[0005] The degree of expansion or dilation in a body vessel such as an artery (e.g. the brachial artery) following blood flow restriction is diminished in patients with CVD and coronary risk factors (e.g. endothelial dysfunction). From a patient compliance perspective brachial artery reactivity as an indicator of CVD is advantageous as it can be measured non-invasively via ultrasound imagery. Expansion or dilation of a vessel, such as the brachial artery, can be estimated by measuring the distance between the arterial walls at one or more cross-sections.

[0006] Estimates of brachial artery reactivity can often be obtained by manually examining either B-mode or M-mode ultrasound images. Manual measurement of arterial diameter is typically done by means of an electronic or virtual calipers built-in to the ultrasound hardware. Manual measurements suffer in that the approach is time-consuming, error prone, and liable to variance depending on the practitioners. Moreover, manual measurements require the presence of the patient during the measurement process. [0007] Techniques for automatically and semiautomatically measuring brachial artery diameter using ultrasound images have also been proposed. For example, Gemignani, et al. ["A system for real-time measurement of the brachial artery diameter in B-mode ultrasound images," IEEE Trans. Med. Imaging, vol. 26, no. 3, pp. 393-404, 2007] discloses a semi-automatic online system for measuring artery diameter based on edge detection and contour tracking. The system requires the practitioner to initialise the procedure by manually tracing two approximate starting borders, which is undesirable as it may introduce a degree of subjectivity and variance in to the process depending on the operative.

[0008] Stadler, et al. ["New methods for arterial diameter measurement from B-mode images," Ultrasound Med. Biol., vol. 22, no. 1 , pp. 25-34, 1996] also communicate a method which focuses on B-mode images of the brachial artery, and the method requires the algorithm to be manually initialised with starting borders.

[0009] Woodman, et al. ["Improved analysis of brachial artery ultrasound using a novel edge- detection software system" J. Appl. Physiol., vol. 91 , no. 2, pp. 929-937, 2001] disclose a semiautomatic system for arterial diameter measurement, in which the system requires three regions of interest to be manually selected by a practitioner, and uses binary thresholding to determine the longest vertical sequences of zeros. This particular system allows for storage and processing of ultrasound images post acquisition.

[0010] Furthermore, the present inventors have found that existing general-purpose

segmentation algorithms, for example, seeded region growing as disclosed by Adams, et al. ["Seeded region growing" IEEE Trans. Pattern Anal. Mach. Intel/., vol. 16, no. 6, pp. 641-647, 1994] perform inadequately for the purpose of segmenting an ultrasound image of an artery in to arterial lumen and arterial tissue. Many general purpose segmentation algorithms are unable to reliably delineate the artery boundary on account of the inherent noise in ultrasound images. In particular, unconstrained region growing algorithms often generate incorrect segmentations, since they do not incorporate prior knowledge about the shape of the region of interest.

[0011] Segmentation algorithms that incorporate prior shape knowledge of tissue structures such as an artery require careful initialisation and significant user interaction in order to achieve reliable results. See, for example, the active contour model of Caselles et al. ["Geodesic active contours," in Proc. 5th International Conference on Computer Vision, 1995, pp. 694-699] and the active shape model of Liu et al. ["Oriented active shape models," IEEE Trans. Med. Imaging, vol. 28, no. 4, pp. 571-584, 2009].

[0012] Unfortunately, most existing segmentation algorithms are too computationally intensive to allow for real-time updates and feedback in a clinical environment.

[0013] Accordingly, there remains a need for an ultrasound image processing tool that is designed to allow medical practitioners to quickly and robustly assess the diameter and wall thickness of arteries from ultrasound images whilst requiring minimal input or interaction from a physician or medical practitioner. Moreover, there remains a need for segmentation methods that facilitate rapid and reliable segmentation of tubular structures such as vascular tissues so as to aid the processing steps discussed above.

Summary of the Invention

[0014] The present invention provides for a method of blood vessel, for example arterial diameter estimation and an associated image segmentation method. The method of arterial diameter estimation disclosed herein is low in cost and can be implemented with minimal investment and development time. The method allows for live feedback and real time adjustment because the processing complexity is greatly reduced relative to prior art methods. Moreover, the method is designed to allow medical practitioners to quickly and robustly assess the diameter and wall thickness of blood vessels such as arteries from ultrasound images.

[0015] The method of the present invention reduces operator subjectivity, and as a result human error, by means of a semi-automatic arterial diameter estimation method. The semi- automated nature of the method disclosed herein makes the method more resilient to noise than manual arterial diameter estimation tools, such as virtual calipers.

[0016] The method of the present invention also provides for post acquisition analysis of ultrasound images, thus, a much greater number of ultrasound images can be recorded at an increased number of time points.

[0017] Accordingly, in a first aspect the present invention provides for a method of processing an ultrasound image of a vascular tissue, the method comprising:

i) segmenting the ultrasound image into regions corresponding to vascular tissue lumen and vascular tissue;

ii) determining the diameter of the vascular tissue lumen at a plurality of points; iii) determining an inter-percentile range of the plurality of diameters; and iv) calculating a statistic of centre of the plurality of diameters from the inter- percentile range of the plurality of diameters.

[0018] As used herein a statistic of centre is a descriptive measure which indicates where the centre or most typical value of a data set lies.

[0019] The step of segmenting the ultrasound image into regions corresponding to vascular tissue lumen and vascular tissue may be done using any segmentation method known to a person skilled in the art. In a preferred embodiment, the step of segmenting the ultrasound image is carried out using the segmentation method of the present invention (vide infra). The ultrasound image may be a 2-Dimensional or 3-Dimensional ultrasound image. For a 2- Dimensional ultrasound image the regions corresponding to vascular tissue lumen and vascular tissue will be represented as vascular tissue lumen pixels and vascular tissue pixels, each different type of pixel having an associated intensity value. For a 3-Dimensional ultrasound image the regions corresponding to vascular tissue lumen and vascular tissue will be represented as vascular tissue lumen voxels and vascular tissue voxels, each different type of voxel having an associated intensity value. As used herein a voxel is a 3-Dimensional volume element in a 3-Dimensional image.

[0020] The method of the present invention is particularly advantageous for processing 2- Dimensional ultrasound images. For example, the ultrasound image may be an M-mode ultrasound image of a vascular tissue. An M-mode ultrasound image is an image that illustrates the positions of moving structures over time. This is particularly advantageous when monitoring cardiac function.

[0021] The method of the present invention may provide for a gating tool which isolates distinct regions of the M-mode scan of a vascular tissue that correspond to scans of the vascular tissue taken at the same point in time in the cardiac cycle. During an ultrasound assessment, patients wear electrodes on their chest to record an electrocardiogram (ECG). Advantageously, a gating tool facilitates diameter assessment specifically at the contraction phase of the cardiac cycle. Consequently, a measurement can be taken at the same point of the cardiac cycle thereby avoiding confounding effects caused by changes in arterial dimensions resulting from the different phases of the cardiac cycle.

[0022] The step of determining the diameter of the vascular tissue lumen at a plurality of points comprises measuring the diameter of the vascular tissue at a given time. The frequency with which these measurements are taken is dependent upon the speed of the ultrasound capture device. The method of the present invention records the diameter of all the spans present in a fixed region of the captured image. For example, the present inventors utilised an ultrasound capture device which had a default setting equating to 450 measurements over a defined time period. The amount of time between these measurements depends on the Ultrasound hardware settings. Varying the capture speed of the ultrasound hardware had no effect on the integrity of the method of the present invention, the latter being robust to changes in the capture time.

[0023] With reference to the method of the present invention the inter-percentile range may be selected from the intertercile range, interquartile range, interquintile range and the intersextile range. Advantageously, an inter-percentile range may help eliminate outliers and imperfections along the tissue lumen boundary, which would otherwise deleteriously affect the accuracy of the method of the present invention.

[0024] In particular, the inter-percentile range may be the interquartile range. The use of the interquartile range may help offset any inaccuracies resulting from the segmentation step. For example, only 50% of the estimated diameters need be considered in determining the statistic of centre. Thus, the scope for error is vastly reduced and the processing power required is cut in half. [0025] The method of the present invention may also provide for highlighting those regions of the vascular tissue lumen that correspond to the range of diameters falling within the inter- percentile range, for example the interquartile range, in the segmented ultrasound image. By highlighting the exact region used in the diameter estimation calculation the practitioner's or user's confidence in the result is greatly increased. Thus, a user can visually, and therefore rapidly inspect that the inter-percentile range of diameters accurately lie within the vascular tissue lumen boundary. This also prevents the practitioner spending time correcting errors that lie outside the inter-percentile range that will have no impact on the final estimation of, for example, the mean arterial diameter.

[0026] For those diameters falling within the inter-percentile range the next step is to calculate a statistic of centre of these diameters. Suitable statistics of centre include at least one of the mean, median, or mode.

[0027] In particular, determining the inter-quartile mean has resulted in diameter estimation that is particularly robust to inaccuracies in the segmentation of the artery.

[0028] The method of the present invention may additionally comprise a plurality of

preprocessing steps. The ultrasound image may be cropped to remove irrelevant information according to the device setup. For example, a pre-smoothing step may optionally be performed to reduce the noise in the image. Prior to the step of segmenting the ultrasound image may be smoothed to remove noise in the ultrasound image. To achieve this the image may be filtered with a Gaussian bilateral filter. Excessive noise in an ultrasound image can diminish the accuracy of medical image processing. Advantageously, preprocessing or smoothing may help reduce noise related inaccuracies without disturbing significant edge information in the image.

[0029] Subsequent to the step of segmenting the ultrasound image, the method of the present invention may further comprise the step of smoothing the image at the boundary of the vascular tissue segments and the vascular tissue lumen segments. This may be done using a 1- Dimensional Gaussian kernel. This may improve the quality and the appearance of the ultrasound image to be processed. This is particularly valuable for an operative of the method of the present invention as it may improve the quality of the automatic diameter measurement.

[0030] The method of the present invention may further comprise providing a caliper tool for measuring the distance between a first selected point and a second selected point in the vascular tissue. This may be particularly advantageous for arterial intima media thickness assessment. For example, the assessment of carotid intima media thickness in the carotid artery.

[0031] The caliper tool allows an operative to click on the innermost point and the outermost point of the intima and media walls and the distance between the selected points is measured. Carotid intima media thickness can be used to monitor cardiovascular disease risk in individuals, monitor disease status, and assess the efficacy of cardiovascular disease treatment. [0032] According to the method of the present invention the vascular tissue may be an artery. For example, the artery may be the brachial artery. The artery may be the carotid artery.

[0033] In a particularly preferred embodiment, the step of segmenting the ultrasound image into regions corresponding to vascular tissue lumen and vascular tissue according to the present invention may comprise applying the segmentation method of the present invention discussed below.

[0034] In a further aspect, the present invention provides for a method of segmenting an image of a tubular structure comprising a hollow core, the image comprising a plurality of image elements and the method comprising:

i) selecting a seed point within the hollow core of the tubular structure and designating the seed point as a candidate image element;

ii) identifying a predetermined number of adjoined image elements horizontally aligned with the candidate image element and designating the candidate image element as a core image element if the predetermined number of adjoined image elements horizontally aligned with the candidate image element have a required intensity value;

iii) selecting a new candidate image element and repeating step ii), wherein iterations of step iii) are performed until a horizontal boundary is reached, such that a horizontal span of the core is provided;

iv) identifying a candidate image element in vertical alignment with a core image element; and

v) designating the candidate image element as a core image element if the candidate image element has a required intensity value and a predetermined number of image elements in (adjacent) horizontal alignment with said core image element (of step iv) are already designated as core image elements.

[0035] The terms horizontal and vertical are to be interpreted in light of the longitudinal axis of the image having the hollow core being segmented. As used herein, the term longitudinal axis refers to the longest axis within the tubular structure/image core.

[0036] Movement parallel to the longitudinal axis passing through the length of the hollow core is designated as the horizontal direction; movement perpendicular to this longitudinal axis is defined as being vertical. For example, where a 2-Dimensional section of a blood vessel is to be segmented the longitudinal axis through the core of the blood vessel will represent the horizontal direction, and all vertical axes will be perpendicular to this longitudinal axis.

[0037] As used herein the term adjoined refers to elements that are touching. References to aligning image elements, aligned image elements and image elements in alignment refer to image elements that would form the points of a straight line.

[0038] Typically, the horizontal expansion stage [i.e., steps i) to iii)] to create a horizontal span of the core takes place before the vertical expansion stage. The seed point is expanded horizontally until it reaches a fixed predetermined boundary. All other expansions are vertical [i.e. steps iv) and v) and optionally steps vi) and vii) below, depending on if the image is 2- Dimensional or 3-Dimensional]. The first series of vertical expansions may create a second horizontal span directly above or below the first horizontal span. The second series of vertical expansions may create a third horizontal span directly above or below the second horizontal span. The segmentation method of the present invention may continue in this manner until the full space is segmented.

[0039] In the segmentation method of the present invention the step of identifying a

predetermined number of adjoined image elements horizontally aligned with the candidate image element may comprise identifying anywhere from 2 to 8 adjoined image elements horizontally aligned with the candidate image element. Typically, the 2 to 8 adjoined image elements horizontally aligned with the candidate image element are on either side (i.e. to the left or to the right) of the candidate image element. In one particular embodiment, this value may be 3 adjoined image elements horizontally aligned with the candidate image element.

[0040] Setting the number high will decrease the tendency for the expanding region to "leak" outside the arterial boundary, but will increase the chances of noise pixels hindering expansion. Setting the value too low will increase the risk of inaccurate leakage of the region boundary and tend to produce non-tubular structures in the presence of gaps in the arterial boundary.

[0041] With reference to the vertical expansion steps, a candidate image element may be added to the set of core image elements if and only if a predetermined number of image elements above or below the candidate image element are already part of the core. Moreover, the image elements above or below the candidate image element must also be in horizontal alignment with the core image element. Preferably, the predetermined number of image elements in horizontal alignment with the core image element is an even number. This enables the predetermined number of image elements to be disposed symmetrically about the core image element.

[0042] Suitable even numbers include 2, 4, and 6. The predetermined value may be 2, i.e. the predetermined number of image elements in horizontal alignment with the core image element in the horizontal span consists of one image element on both sides of said core image element in the horizontal span. This value allows for rapid processing of the method of the present invention.

[0043] Setting the number high will decrease the tendency for the expanding region to "leak" outside the arterial boundary, but will increase the chances of noise pixels hindering expansion. Setting the value too low will increase the risk of inaccurate leakage of the region boundary and tend to produce non-tubular structures in the presence of gaps in the arterial boundary.

[0044] Advantageously, the segmentation method of the present invention is not

computationally intensive and may allow for rapid modification of threshold intensity values/parameters and real-time updating of the segmentation in response to such modifications to the threshold intensity values/parameters. For example, the threshold of the required intensity value for a core pixel can be changed. Alternatively, a new seed point can be selected such that the segmentation can be recalculated in real-time.

[0045] With reference to 2-Dimensional ultrasound images, such as M-mode images, the images are composed of an array of pixels that are individually displayed as a function of differences in tissue densities resulting from the refraction or transmission of high-frequency sound waves. Each pixel is composed of 8-bits (or 1 byte). Thus, each pixel corresponds to 8 binary bits and may assume anyone of 256 values depending on the combination of zeros and ones adopted by the 8 bits. The particular combination of ones and zeros in each 8-bit pixel is proportional to the density of the tissue which the pixel represents (which is proportional to the frequency of the returned sound wave by that tissue). For example, using a grey scale colour system, a pixel may take a value from 0 indicating black to 255 indicating white, with values therebetween corresponding to varying shades of grey.

[0046] As used herein, for a 2-Dimensional M-mode ultrasound image, pixel intensity is utilised to indicate a pixel having one of the 256 greyscale values from 0 to 255. For example, 0 intensity corresponds to a black pixel whereas 255 intensity corresponds to a white pixel. With reference to the segmentation method of the present invention, the default threshold for expansion in the vertical direction may be set to an intensity of 30. If the intensity of a pixel is below this value, it is not added to the current front when expanding in the vertical direction.

[0047] As will be appreciated by a person skilled in the art, the display of the ultrasound image is achieved by means of a mapping function which maps a particular pixel intensity (from 0 to 255) to a particular luminance (cd/m²) for the on-screen display.

[0048] Advantageously, the segmentation method of the present invention allows the threshold values/required intensity values to be changed thereby allowing rapid re-segmentation of the image to suit the needs of an operative/user. Thus, if a user is not happy with the results of the initial segmentation the threshold intensity value can be adjusted to allow a real-time, near instantaneous re-segmentation of the image based on the new threshold intensity value. This allows a user to rapidly assess the accuracy of a segmentation at a particular threshold intensity and to modify or improve the segmentation by altering that threshold intensity value.

[0049] With reference to the segmentation method of the present invention the image element may be a 2-Dimensional image element, such as a pixel. Similarly, the image element may be a 3-Dimensional volume element, such as a voxel. As will be appreciated by a skilled person, where the image elements are voxels or volume elements and the image is a 3-Dimensional image, the method of the present invention may further comprise iterations of the following steps: vi) identifying a candidate volume element in linear (or coaxial) alignment with a core volume element, wherein the core volume element is in a plane of volume elements formed by steps i) to v); and

vii) designating the candidate volume element as a core volume element if the candidate volume element has a required intensity value, and a predetermined number of volume elements in either vertical or horizontal (adjacent) alignment with said core volume element are already designated as core volume elements.

[0050] The term linear (or coaxial) alignment refers to positioning the two image elements such that one sits directly over the other, i.e. they superpose or eclipse one another. Essentially, the candidate volume element of step vi) is perpendicular to the plane of volume elements formed by steps i) to v) and it superposes or eclipses the core volume element of said plane.

[0051] In step (vii) the predetermined number of volume elements in either vertical or horizontal alignment with the core volume element may comprise an even number. This may allow for a symmetric disposition about the candidate volume element. For example, 2 to 8. Suitably, the even number may be 2.

[0052] The segmentation method of the present invention may be applied to a tubular structure, such as a vascular tissue. The vascular tissue may be selected from an artery and a vein. The artery may be the brachial artery. The artery may be the carotid artery. Accurate and robust segmentation of such structures may be particularly advantageous in medical image

processing.

[0053] In a particularly preferred embodiment, the ultrasound image processing method of the present invention is combined with the segmentation method of the present invention.

[0054] In yet a further aspect the present invention provides for a computer readable medium comprising program instructions which when executed by a processor perform:

the ultrasound image processing method of the present invention; and/or

the segmentation method of the present invention.

[0055] The present invention also provides for a computer implemented system for processing an ultrasound image of a vascular tissue, the system comprising:

i) a segmenting tool for segmenting the ultrasound image into regions

corresponding to vascular tissue lumen and vascular tissue;

ii) a processing means for determining the diameter of the vascular tissue lumen at a plurality of points;

iii) a processing means for determining an inter-percentile range of the plurality of diameters; and

iv) a processing means for calculating a statistic of centre of the plurality of diameters from the inter-percentile range of the plurality of diameters. [0056] The present invention also provides for a method of detecting cardiovascular disease in a patient comprising:

i) providing an ultrasound image of a vascular tissue of the patient; and

ii) applying the ultrasound image processing method of the present invention to the image.

[0057] The present invention also provides for a method of detecting endothelial dysfunction in a patient comprising:

i) providing an ultrasound image of a vascular tissue of the patient; and

[0058] Endothelial dysfunction is an indicator in the onset of cardiovascular disease.

[0059] The method of the present invention can be used to process an ultrasound image of a vascular tissue for the detection of cardiovascular disease.

[0060] Where suitable, it will be appreciated that all optional and/or preferred features of one embodiment of the invention may be combined with optional and/or preferred features of another/other embodiment(s) of the invention.

Brief Description of the Drawings

[0061] Additional features and advantages of the present invention are described in, and will be apparent from, the detailed description of the invention and from the drawings in which:

[0062] Figure 1 shows a screenshot of a user interface for performing the method of the present invention;

[0063] Figure 2 shows a flow chart illustrating the major processing steps of the method of the present invention;

[0064] Figure 3 illustrates the segmentation method of the present invention;

[0065] Figure 4 illustrates the contraction phase in an ECG; and

[0066] Figures 5 and 6 illustrate scatterplots showing the gated and ungated measurements of the present invention plotted against manual measurements taken by the sonographer

Detailed Description of the Drawings

[0067] It should be readily apparent to one of ordinary skill in the art that the examples disclosed herein below represent generalised examples only, and that other arrangements and methods capable of reproducing the invention are possible and are embraced by the present invention.

[0068] Figure 1 shows a screenshot of a user interface for performing the method of the present invention. To use the tool for diameter assessment, a practitioner first selects one or more ultrasound images 101 of a large artery to analyse. For each image, the practitioner locates the artery and clicks the area 102 between the arterial walls. The method of the present invention uses the point provided by the practitioner to segment the arterial boundary using a constrained region-growing algorithm (as discussed below), and the result is depicted visually in that the segmented arterial lumen is highlighted using grey shading 103. If the practitioner is not happy with the segmentation he or she may change the threshold intensity values of the segmentation method using a slider 104, or pick a new seed point and start again. The segmentation of the artery is updated in real-time and the user can increase or decrease the threshold intensity values using the slider 104.

[0069] An estimate of the arterial diameter is computed from the segmentation using a subset of the vertical diameters that correspond to the inter-quartile range. The cross-sections of the artery that correspond to the diameters within the interquartile range may be shown in a different or lighter colour so that the practitioner may easily distinguish them. This may be achieved using different shades of grey in the highlighted area 103 between the arterial walls. This allows a practitioner to visually inspect the segmentation, thus, inspiring greater confidence that the segmentation is accurate. Thus, for the interquartile range, a practitioner need only examine that the 50% of the arterial lumen corresponding to the diameters within the interquartile range has been properly segmented because calculations are only done on this 50%. This strategy ensures that outliers and imperfections along the boundary have minimal effect on the final estimate, while also ensuring that practitioners do not spend time making corrections that do not affect the estimate. The method of the present invention also supports gated measurements, i.e. estimates taken at specified vertical cross-sections, and software calipers for manual measurements.

[0070] Figure 2 shows a flow chart illustrating the major processing steps of the method of the present invention. Processing is divided into the following stages.

[0071] Image acquisition: The practitioner selects one or more images for analysis using a standard dialog box. If the ultrasound source is a video sequence, the still frames of interest must first be extracted prior to using the tool. The software can handle images stored in various file formats including BMP, PNG, JPEG, and GIF.

[0072] Calibration: A device specific calibration step establishes the number of pixels per centimetre in each ultrasound image. The implementation is based on a Sonosite MicroMaxx ® ultrasound system. Calibrating this device requires first detecting the mode, zoom level, and depth, followed by counting the number of pixels between the measurement bars on the right- hand-side of the image. The mode and zoom-level are recognized by template matching at predefined areas of the image. The depth is recognized by template matching using prototypes of the digits zero through nine in the bottom right hand corner of the image (for example, in Figure 1 the depth is 4.0). The distance between the measurement bars in centimetres is a function of the mode, zoom-level, and depth. As such, counting the number of pixels between each measurement bar may be used to determine the number of pixels per centimetre in the image.

[0073] Preprocessing: The image is cropped to remove irrelevant information according to the device setup. Excessive noise can diminish the effectiveness of medical image processing algorithms. As such, a pre-smoothing step is optionally performed to reduce the noise in the image. In this step, the image is filtered with a Gaussian bilateral filter. This helps ameliorate noise without disturbing significant edge information in the image. Preprocessing can be disabled in the user interface if desired.

[0074] Segmentation: The practitioner identifies the artery of interest, and notifies the application of its position by clicking inside it. A constrained seeded region-growing algorithm is used to delineate the arterial boundary based on the position indicated by the practitioner and a live adjustable threshold. This algorithm is described in detail below.

[0075] Post-processing: The region boundary is smoothed by convolving it with a 1 -dimensional Gaussian kernel.

[0076] Robust statistics: The height of each vertical span in the segmentation is calculated, and the inter-quartile range of these values found. The mean of this range is the diameter estimate. The spans that contribute to the estimate are highlighted as feedback to the practitioner, and the estimate displayed.

[0077] Refinement: The practitioner observes the feedback and diameter estimate from the previous step and decides if the estimate is satisfactory. At this stage, the practitioner may choose to adjust the threshold used in the segmentation step with a slider 104. As the slider 104 is moved the feedback and diameter estimate are updated in real time (see Figure 1).

[0078] Additional tools: The application includes built-in software calipers that allow manual measurements to be taken at fixed points in conjunction with the diameter estimates. A gating tool is included to allow measurements to be taken at specific cross-sections of the estimated arterial region.

Segmentation method

[0079] The segmentation method of the present invention was designed to satisfy the following requirements:

[0080] It must be fast enough to allow for updating the seed points and parameters while providing real-time feedback to an operator.

[0081] It must be able to reliably achieve a reasonably accurate segmentation. More precisely, it is required to be able to achieve near perfect accuracy for at least 50% of the vertical pixel spans in the segmentation after a suitable adjustment of the seed point and parameters.

[0082] The segmentation method operates in two stages: horizontal expansion, and vertical expansion. In the horizontal expansion stage the seed point is expanded to a single pixel thick horizontal connected path. This is achieved by growing the seed point in both the left and right directions using a simple greedy strategy. This strategy begins by setting the current pixel to the seed pixel. If we are growing the line in the right direction, at each step, the three neighboring pixels to the right of the current pixel are examined. The current pixel is updated to be the darkest of these and is added to the line. This process continues until the expansion has reached a predefined boundary.

[0083] In the vertical expansion stage the region is grown upwards and downwards using a dynamic programming approach. The region is expanded upwards as follows. A series of 3 pixels designated, for example as core region pixels are identified. These are labelled XXX in Figure 3. A candidate pixel, i.e. pixel C in Figure 3, is identified. The candidate pixel C is in vertical alignment with the middle pixel of the core pixels labelled XXX in Figure 3. Thus, the pixels XXX are disposed symmetrically using candidate pixel C as a centre point. For the sake of completeness, the bottom row of shaded pixels equates to the first horizontal span of the core region of the image. If the grey scale intensity of candidate pixel C corresponds to a certain threshold, and the three neighbouring pixels (i.e. pixels labelled XXX) below it are already part of the core region, then the candidate pixel C is added to the core region. This process is repeated recursively until the core region is fully expanded. A similar strategy is followed to expand the region downwards.

Diameter estimation system evaluation

[0084] The method of the present invention was evaluated by comparing estimates taken with the system to manual measurements taken by a trained sonographer with a virtual calipers. The manual measurements were taken with a Sonosite MicroMaxx ® ultrasound system, specifically at the contraction phase of the cardiac cycle (QRS gating). The contraction phase as seen in an ECG is illustrated by the grey highlighting in Figure 4.

[0085] These gated measurements are directly comparable to the gated estimates taken at the same contraction phases with the method of the present invention. The correspondence between the manual measurements and both the gated and ungated estimates produced by the method of the present invention are discussed below.

[0086] The method of the present invention was evaluated using 8 subjects. The sonographer took a total of 144 measurements of brachial diameter. To measure the agreement between the manual expert measurements and the measurements produced by the method of the present invention, the Pearson product-moment correlation coefficient was utilised. The coefficient measures the linear correspondence between two variables, that is, how well one variable could be approximated as a linear function of another variable. The correlation coefficient lies in the range [-1 , 1], where 1 indicates perfect linear correspondence, -1 indicates perfect negative linear correspondence, and zero indicates that there is no linear correspondence. [0087] For the gated measurements, the correlation coefficient between the values produced by the method of the present invention and those produced by the sonographer was 0.988, indicating a near perfect linear correspondence. This correlation value is specific to the experiment performed. Assuming measurements follow a normal distribution, the true statistical correlation can be shown to be between 0.98 and 0.99 with 95% confidence.

[0088] For the ungated estimate, produced by taking the interquartile mean of the diameters in all vertical spans in the segmented area, the correlation coefficient was 0.927, again showing a very high linear correspondence. As there is more variance in the measurements, the confidence interval is a little larger, and the true statistical correlation can be said to be between 0.90 and 0.95 with 95% confidence.

[0089] Of course, since the manual measurements and the gated measurements only apply to the contraction phase of the cardiac cycle, and the interquartile mean is taken across one or more cycles, the correlation between the ungated estimates is expected to be lower than that of the gated estimates.

[0090] Figures 5 and 6 illustrate scatterplots showing the gated and ungated measurements of the present invention plotted against manual measurements taken by the sonographer. The sonographer's measurements are plotted along the vertical axis; the measurements produced by the method of the present invention are plotted along the horizontal axes. The central line represents the linear regression line. The gated measurements are shown in Figure 5. The interquartile mean (IQM) based estimate in Figure 6. The plots demonstrate that there is high correlation between the measurements produced by the method of the present invention and the measurements of an expert sonographer.

[0091] In addition to comparing measurements of the present invention against those of a sonographer, the accuracy of the segmentation method was also evaluated by comparing segmentations created by the method with a manually constructed ground truth. The test set comprises 65 M-Mode images of the brachial artery taken from 3 subjects. To create the ground truth, the region between the arterial walls was manually outlined using image-editing software. The segmentation method of the present invention was then used to interactively segment the interior of the arteries from the same images and the results were compared using two symmetric accuracy metrics: object accuracy and boundary accuracy.

[0092] Object accuracy measures the degree of overlap between the object pixels in the ground truth (G) and those in the machine segmentation (M), where the object pixels are those marked as being inside the arterial walls. Its value is equal to the Jaccard index: the ratio of the size of the intersection of M and G to the size of the union of M and G:

[0093] Boundary accuracy is a fuzzy version of the Jaccard index, and measures the accuracy of the object boundary. Both measures give values that lie in the interval [0, 1], where higher values indicate greater accuracy. A more thorough description of these measures is given in McGuinness & O'Connor [A comparative evaluation of interactive segmentation algorithms, Pattern Recognition, vol. 43(2), 2010, pp. 434 - 444.]

[0094] Using these two measures the mean object accuracy was 0.955±0.006 and the mean boundary accuracy was 0.869±0.017. The confidence intervals were computed based on an assumed t-distribution for a 95% confidence level. Interactively segmenting all 65 images using the method of the present invention took a total of 6 minutes: around 5.5 seconds per image. Table 1 summarizes the accuracy values.

Table 1

[0095] The mean values compare favourably with accuracy measurements from user experiments using general purpose interactive segmentation algorithms on natural images [McGuinness & O'Connor, 2010, supra], wherein mean accuracy values for state-of-the-art interactive segmentation algorithms were reported to be 0.92 and 0.78 for object and boundary accuracy. The segmentation method of the present invention gives higher accuracy and requires less interaction.

[0096] The words "comprises/comprising" and the words "having/including" when used herein with reference to the present invention are used to specify the presence of stated features, integers, steps or components but do not preclude the presence or addition of one or more other features, integers, steps, components or groups thereof.

[0097] It is appreciated that certain features of the invention, which are, for clarity, described in the context of separate embodiments, may also be provided in combination in a single embodiment. Conversely, various features of the invention which are, for brevity, described in the context of a single embodiment, may also be provided separately or in any suitable subcombination.

Claims

1. A method of segmenting an image of a tubular structure comprising a hollow core, the method comprising the steps of:

ii) identifying a predetermined number of adjoined image elements horizontally aligned with the candidate image element and designating the candidate image element as a core image element if the predetermined number of adjoining image elements horizontally aligned with the candidate image element have a required intensity value; iii) selecting a new candidate image element and repeating step ii), wherein iterations of step iii) are performed until a horizontal boundary is reached, such that a horizontal span of the core is provided;

v) designating the candidate image element as a core image element if the candidate image element has a required intensity value and a predetermined number of image elements in horizontal alignment with said core image element of step iv) are designated as core image elements.

2. A method according to Claim 1 , wherein the step of identifying a predetermined number of adjoined image elements horizontally aligned with the candidate image element may comprise identifying from 2 to 8 adjoined image elements horizontally aligned with the candidate image element.

3. A method according to Claim 2, wherein the 2 to 8 adjoining image elements horizontally aligned with the candidate image element are on either side of the candidate image element.

4. A method according to any one of Claims 1 to 3, wherein in step (v) the predetermined

number of image elements in horizontal alignment with the core image element comprises an even number.

5. A method according to any one of Claims 1 to 4, wherein the required intensity value is changeable.

6. A method according to any one of Claims 1 to 5, wherein the tubular structure is a vascular tissue.

7. A method according to Claim 6 wherein the vascular tissue is selected from an artery and a vein.

8. A method according to any one of Claims 1 to 7 wherein the image element is a 2- Dimensional pixel or a 3-Dimensional volume element.

9. A method according to Claim 8 wherein where the image element is a 3-Dimensional volume element and the method further comprises the following steps:

vi) identifying a candidate volume element in linear alignment with a core volume element, wherein the core volume element is in a plane of volume elements formed by steps i) to v); and

vii) designating the candidate volume element as a core volume element if the

candidate volume element has a required intensity value, and a predetermined number of volume elements in either vertical or horizontal alignment with said core volume element are designated as core volume elements.

10. A method according to Claim 9, wherein in step (vii) the predetermined number of volume elements in either vertical or horizontal alignment with the core volume element comprises an even number.

1 1. A method according to Claim 6, wherein the image is segmented into regions corresponding to vascular tissue lumen and vascular tissue, the method further comprising the processing steps of:

(a) determining the diameter of the vascular tissue lumen at a plurality of points;

(b) determining an inter-percentile range of the plurality of diameters; and

(c) calculating a statistic of centre of the plurality of diameters from the inter- percentile range of the plurality of diameters.

12. A method of processing an ultrasound image of a vascular tissue, the method comprising:

13. A method according to Claim 12 wherein the inter-percentile range is selected from the intertercile range, interquartile range, interquintile range and the intersextile range.

14. A method according to Claim 13 wherein the inter-percentile range is the interquartile range.

15. A method according to any one of Claims 12 to 14, wherein the statistic of centre is at least one of the mean, median, or mode.

16. A method according to any one of Claims 12 to 15, wherein prior to the step of segmenting the ultrasound image is smoothed to remove noise in the ultrasound image.

17. A method according to any one of Claims 12 to 16, wherein subsequent to the step of segmenting the ultrasound image is smoothed at a boundary of the vascular tissue and the vascular tissue lumen.

18. A method according to any one of Claims 12 to 17, wherein the ultrasound image is an M- mode ultrasound image of a vascular tissue.

19. A method according to Claim 18 further comprising providing a gating tool for isolating distinct regions of the M-mode scan of a vascular tissue that correspond to scans of the vascular tissue taken at a same point in time in a cardiac cycle.

20. A method according to any one of Claims 12 to 19, further comprising providing a caliper tool for measuring the distance between a first selected point and a second selected point in the vascular tissue.

21. A method according to any one of Claims 12 to 20, wherein the step of segmenting the ultrasound image into regions corresponding to vascular tissue lumen and vascular tissue comprises applying a method according to any one of Claims 1 to 10.

22. A method according to any one of Claims 12 to 21 , wherein the vascular tissue is an artery.

23. A computer readable medium comprising program instructions which when executed by a processor perform the segmentation method of any one of Claims 1 to 1 1.

24. A computer readable medium comprising program instructions which when executed by a processor perform the processing method of any one of Claims 12 to 22.

25. A computer implemented system for processing an ultrasound image of a vascular tissue, the system comprising:

i) a segmenting tool for segmenting the ultrasound image into regions corresponding to vascular tissue lumen and vascular tissue;

iv) a processing means for calculating a statistic of centre of the plurality of diameters from the inter-percentile range of the plurality of diameters.

26. A method of detecting cardiovascular disease in a patient comprising:

i) providing an ultrasound image of a vascular tissue of the patient; and ii) applying the method of any one of Claims 12 to 22 to the image.

27. Use of the method of any one of Claims 12 to 22 in processing an ultrasound image of a vascular tissue for the detection of cardiovascular disease.